AMD is already at the top of the CPU market. It was second once to Intel. Today the only excuse i hear about buying Nvidia GPUs is CUDA. What about ROCm?
amd was always trash
https://www.amd.com/en/products/software/rocm.htmlThe AMD ROCm page has several references to CUDA but the NVIDIA CUDA page makes no mention of ROCmhttps://developer.nvidia.com/cuda
>>110281320Intel used to make no references to AMD either, didn’t help them much in the long run. Only thing that matters is good products with a consistent schedule.
>>110281320Yes, AMD is the one that has to rise to the level of CUDA, the question is, is it already there? Radeon Gpus have better value.
>>110281342>Intelwho?
>>110281349>Radeon Gpus have better value.To who?
>>110281273I was doing some graphics programming with glsl, and AMDs "texture quality" setting destroyed my anit-aliasing. Making every pixel more contrasty, and fucking with the image. It is this stupid place, where we developers want to get a good image, the gpu-driver has a different opinion on what a good image is, and there is slop at the middle. Does anybody know how to turn this shit off for the end user in sdl3??
>>110281273ROCm is dead last.1st: CUDA2nd: MLX (Apple)3rd: OpenVINO (Intel)4th: QAIRT (Qualcomm) (SD8 G3 and later can run SDXL directly on the Tensor Cores)5th: CPU6th: ROCmAMD is still big because they pretty much dominate the CPU space. Ryzen scales extremely well. Apple might be good with single core, but they don't sell servers.
>>110281273>What about ROCm?I'm finding it usable for image/video gen with a Ryzen iGPU. A few specific pain points have improved over the last year or two after my hardware got ROCm support.>ComfyUI now disables pinned memory for iGPUs by default, since it's pointless for them. This avoids a freezing issue with Ryzen if RAM usage maxes out, and the underlying ROCm bug was also tracked down and fixed recently by AMD, though not released yet.>VAE decode (the last step of image gen) is mostly good now if you set an env var for COMFYUI_ENABLE_MIOPEN = 1. MIOpen used to run like shit, so ComfyUI disabled it by default. ROCm 7.14 fixed the perf issues for me, although 10.1 has regressed a bit while AMD fiddles with the first-run benching behavior. It's still helpful overall since it greatly reduces VAE decode's RAM usage.>ComfyUI introduced Comfy Kitchen attention, which is competitive with flash/sage attention while being way easier to enable on AMD than those.>I realized running Comfy with --disable-smart-memory lets me run bigger models/gens without maxing RAM
Is this a system-seller??
I've used CUDA, RocM and IPEX and out of all three of them RocM is the most painful.CUDA just works, you can have a 2GB gtx 1050 and it will work with the last CUDA version before Nvidia dropped support for Pascal. On the other hand my RDNA 2 cards work like half the time, I often have to export environment values to get them to work and sometimes they don't even work at all (Or worse, they work once and then they shit the bed and I must spend the entire afternoon troubleshooting shit).AMD was late to this race and once they entered it they still dragged their feet for a while regarding RocM compatibility for consumer cards, to the point where honestly using Vulkan is better when given the choice for their older cards, even Intel got the memo and they made their APIs available from day one and they still work for all Arc models except the ones with 4GB of vram, I cannot say the same for AMD; I have a 16GB Polaris card and I cannot even get RocM to work on it.
>>110281891>>110281735I was wondering about what experience ppl had with ROCm and if it was good or not. For what you say is kind of crap.
>>110281365If you look at their speed and memory they are quite competitive for the price. But if you cant use CUDA and have to stick with ROCm then is not that good
>>110281273>What about ROCm?CUDA vs. ROCm has become less and less relevant in the past few years after the ai hypetrain finally lit a fire under AMD's ass to get their compute stack up to par. I'd say its almost there, but it's also not the equivalent to practically unusable garbage like it used to be. I think RDNA5 might be their Zen moment where they finally design a GPU for not just gaming performance, but also feature parity for everything else
>>110281273AMDs ecosystems are largely open source. It's only when you get FOMO gam3rs and BIZnsss people get you nihilistic edgelord cicrcljerking about how AMD is losing on every front.
>>110281273I think AMD graphics cards are objectively better for vidya.
